Happy Halloween. The scariest thing in tech this week wasn't a costume. It was version 2.1.313 of a coding extension, published under the right account and delivered by auto-update. Trick or treat, and the extension picked trick. This is AI Daily Briefing, and it's Saturday, October 31st. Today: a Claude Code extension on Cursor's marketplace gets backdoored, Thrive Holdings and OpenAI go shopping for a home security company, Deepgram buys the team behind the Pipecat voice framework, and Fujifilm bets the AI boom needs a lot more magnetic tape.
Dan Goodin, writing in Ars Technica:
Cursor said that the malicious extension attempts to install a persistent SSH backdoor using an added attacker-controlled SSH key, enabling the local SSH service, and opening a reverse SSH tunnel. The code was written to work on Windows, macOS, and Linux.
Here's the sequence. Cursor said Friday that version 2.1.313 of the Claude Code extension in its marketplace was malicious and reached at most five thousand users. Ars puts the exposure window at nearly twelve hours. Independent researcher Kevin Beaumont and Luke Roberts, the CTO and co-founder of the security firm Phorian, both say it went out through Anthropic's official account on Cursor. Anthropic pushed 2.1.314 several hours later. And that's the scary part. Nobody typo-squatted a lookalike package. Roberts told Ars the attacker got into the official distribution channel, and the bad version simply superseded the good one. Extensions in Cursor auto-update by default. So the people who did everything right, who installed the real extension from the real publisher, were the ones who got hit. And think about what an extension can reach. Ars notes Cursor extensions run with the permissions of the IDE itself, so the backdoored build likely had access to local files, source code and developer tooling. On a developer laptop that usually means keys, tokens and every repo the person touches. The coding agent isn't what made this dangerous. The danger is where it lives. We give these tools the terminal, the repo and the credentials because that's what makes them useful. That's also the blast radius. Two open questions. First, neither Cursor nor Anthropic has explained how the attackers got control of that account. Second, as of the Ars piece, Anthropic had yet to issue public notice to customers. The 12PB Harbor digest of Cursor's disclosure says Cursor took its extension marketplace offline that afternoon, cleaned it up with the providers involved, and is telling affected users to check their systems and move to 2.1.314. My Monday-morning list: find out who on your team runs Cursor, check which extension version they had during the window, and seriously consider turning off auto-update for anything that holds credentials. Pin versions for extensions the same way you pin dependencies.
Julie Bort, writing in TechCrunch:
Thrive Holdings is the holding company spinout from Josh Kushner’s powerhouse VC firm Thrive Capital. It also counts OpenAI as its partner, as well as an investor. The firm’s mission is to hold other companies long-term, revitalizing them with AI.
The target is ADT. The companies announced Thursday a deal at ten dollars a share, about seven point six billion in equity and fifteen point three billion in enterprise value, by ADT's count. AI Weekly's summary adds a forty-five point six percent premium, a thirty-day go-shop window, and Apollo financing with no financing contingency. So technically this is a signed deal, not a closed one, despite the headline. The OpenAI part is the interesting bit. TechCrunch says OpenAI supplies Thrive's companies with models, tech and forward-deployed engineers. Thrive says it already holds more than seventy-five companies. And AI Weekly quotes OpenAI COO Brad Lightcap calling the deal a template for frontier AI deployed rapidly across entire organizations. Notice what isn't in the announcement. Neither company named a single new AI product. The press release only says ADT will form the foundation of Thrive Holdings' smart home security business. ADT already partners with Apple and Google, and TechCrunch expects that to continue. So the bet is that you can buy a big, slow-moving services company, put forward-deployed engineers inside it, and make the margins look like software. Maybe. But a fifteen-billion-dollar enterprise value is a big price for a thesis with no product attached yet. I want to see what an AI-rebuilt ADT actually ships before I call it a template. It isn't the only buyer like that this week. Bort also reported Friday that Crosby, a two-year-old AI-first law firm, bid for some of Weil, Gotshal and Manges's assets, including the trademark, the brand and some internal data, as that firm prepares to vote on dissolution. Crosby is also trying to hire some of Weil's twelve hundred lawyers. Terms weren't disclosed, and there's no guarantee it wins. A startup that has raised about eighty-five million dollars, bidding on a Wall Street brand name. The AI-native-services crowd isn't just disrupting incumbents anymore. They're buying them.
Ivan Mehta, writing in TechCrunch, quoting Daily co-founder Kwindla Kramer:
Scott was super committed to Pipecat being open source and vendor-neutral forever. I think there is both a personal commitment and commercial driver there.
Deepgram said Thursday it's acquiring Daily, the maker and maintainer of Pipecat, the open-source framework that wires together speech-to-text, the language model and text-to-speech in a real-time pipeline. No price disclosed. Both companies say Pipecat stays open source. Daily has raised over sixty-two million dollars, per PitchBook, and Deepgram hit unicorn status earlier this year with a hundred-and-thirty-million-dollar Series C. They were in the same Y Combinator batch, back in 2016. I've built on this stack, and it's where voice agents live or die. Your model can be great, but if the turn-taking latency is bad, the call feels broken. So a model company buying the orchestration layer makes sense. Vendor-neutral forever is the promise to watch. CEO Scott Stephenson told TechCrunch he wants developers using the best model for their use case, even when that's not Deepgram's. They're also launching what they call the Real-Time Context Standard, an open protocol co-developed with the Pipecat team. It carries signals from the audio, say, that the user is whispering, plus instructions for the language model on how to use those signals and guidance on how the response should sound. A standard from one vendor is a proposal until competitors adopt it. If LiveKit or Vapi pick it up, that's news. Until then, it's Deepgram's spec. And the model side keeps getting crowded. The AIBriefs roundup today lists Nvidia's Nemotron 3.5 Transcribe, an open-weight thirty-billion-parameter speech model that leads Hugging Face's Open ASR Leaderboard among open models at four point seven one percent mean word error rate. One caveat: it works only on recordings, not live audio. Which means it's not dropping into a Pipecat voice agent anytime soon. Great for transcripts. Not for a live phone call.
Ars Technica, on a Fujifilm announcement:
In an announcement released Thursday, Fujifilm said it will invest roughly 10 billion yen (about $63 million) to increase its production capacity fivefold for an unlikely beneficiary of the AI era: magnetic tape.
The fivefold increase is measured against fiscal 2025 output. The money goes into Fujifilm's Odawara site in Kanagawa, runs from fiscal 2026 through 2028, and mainly produces LTO Ultrium 10 cartridges. Fujifilm calls it the largest investment it has made in dedicated LTO tape facilities. Its explanation is that generative AI is driving what it calls exponential growth in data. Every training run, every checkpoint and every eval log has to live somewhere. Not all of it needs to sit on flash. Fujifilm says major cloud and data center operators are expanding their tape use because it's cheap at volume. There's a security angle too, which loops back to our lead. Tape can be physically disconnected and air-gapped, and Fujifilm pitches it as a last line of defense against ransomware. Sixty-three million dollars is a rounding error next to GPU capex. But it's a real capacity number, from a supplier nobody thinks about, and that's the kind of signal I trust more than a keynote. AI demand is pulling in stranger suppliers too. Ars also reports that Boom Supersonic, the company building the Overture supersonic airliner, has shifted much of its focus to natural gas turbines for data centers. On October 22nd it unveiled a factory in Centennial, Colorado, to build a forty-two-megawatt turbine it calls Superpower. Its Overture factory in Greensboro, North Carolina has mostly sat dormant since it was finished in 2024. The plane is still waiting on certification, but data centers are buying turbines now.
Did the extension story send you off to audit your team's IDEs this weekend? Tell us what you found, or what you'd want us to dig into next. Write to aidailybriefing at lantern podcasts dot com. We read every message.
One thing to do before the candy runs out: check your extension versions. And watch for Anthropic's own disclosure on how that account was compromised. AI Daily Briefing is part of the Lantern Podcasts family. Our next episode is Monday, November 2nd.